multi-hop-rag

Multi-hop retrieval chains retrievals across documents to synthesize evidence. It decomposes your question into steps, retrieves relevant passages iteratively, and synthesizes a coherent answer with cited sources—essential for cross-document reasoning in research, legal, and enterprise contexts.

Standard RAG retrieves once; multi-hop RAG performs iterative retrievals, refining results between steps to connect facts across sources. This approach handles complex questions requiring reasoning across multiple documents that single-pass retrieval cannot answer completely.

You need a vector store for document embeddings, an LLM for reasoning and synthesis, and orchestration logic to decompose questions and chain retrieval steps. The vector store indexes your documents; the LLM manages multi-step reasoning.

Yes. Multi-hop RAG uses vector search to retrieve passages across hops and integrates with vector stores like Milvus for scalable document indexing and retrieval across multiple steps.

Multi-hop RAG returns synthesized answers with cited sources and context from each retrieval hop, making the reasoning path transparent. You can see which documents contributed to each step and how facts connected across sources.

Use multi-hop retrieval when questions require connecting facts across multiple documents, iterative discovery, or reasoning that single-pass retrieval cannot handle. It's suited for complex research, legal analysis, and enterprise knowledge synthesis.